Business, Technology

No matter the levels of automation nor intelligence – your tools always need a human touch

Four human characters operating a working humanoid-robot, which combines the intelligence of machines and humans.
Illustration by Reetta Kotilainen.

AI solutions, even built by the best data scientists, will be worthless investments unless they are developed and implemented with a multidisciplinary team. Pouring in money to obtain high autonomy or so-called intelligence does not remove the need for versatile skills and perspectives when aiming for a profitable solution. It just makes it more difficult and too many wonderful AI systems have failed because of poor UX or simply internal marketing.

In this blog post, we discuss the importance of multitalented teams when creating smart solutions to real-world problems. Also, with examples from sensor fusion, we demonstrate why neither the autonomy nor the complexity of a solution guarantees its profitability.

Henrik Aalto is a business-savvy data scientist with a strong background in operations research. At Reaktor, he’s honing in on delivering measurable value to clients.

Ville Rantanen is a data scientist, who loves to spread the understanding of technology to non-experts. He specializes in machine learning and machine vision.

Regardless of the industry, utilizing smart solutions is nowadays essential to stay competitive. Some very special domains aside, the competitors will anyways put new technological advancements into use and merely keeping up to them, as the Red Queen once said to Alice in Through the Looking-Glass, takes all the running you can do. And obviously, taking an edge over one’s competitors takes even more than that. But how much?

Although some smart digitalization is essential across businesses, the degree to which it is needed as of today depends on the margins. The narrower they are, the broader and deeper the minimum degree is. And today, smart solutions are ubiquitous right below the surface layer. “But my business is making a good profit, why should I invest?” one might ask. The answer is: the margins discussed here must not necessarily be economic. A currently profitable business can soon go down if the technological competition is fierce and a company cannot keep up with its rivals. Be the requirements economic, technological, or something else, modern technology is often useful in meeting them.

Favorite taste of AI: Your success will only come with a multidisciplinary approach

One of the first decision points a leader faces after deciding to invest in technology is the scope. If picking one’s favorite flavor from 24 varieties of jam is nearly impossible, so must be assessing the huge assortment of solution possibilities ranging from zero-capex SaaS products to one-stop-shop custom software providers and even developing everything in-house, which would serve as its topic. What makes the decision more challenging is the fact that proper scoping might make the difference between a hugely successful investment and releasing a hydra inside a company’s IT infrastructure.

When we think of smart solutions, (or, advanced analytics, AI, intelligent tools), it is perhaps more important to consider the levels of autonomy and intelligence of the system than the price tag or the question of insourcing vs. outsourcing. Thinking of the autonomy and intelligence dimensions as a fourfold table, the archetypes can be named as simple tools, augmented intelligence, classic automation, and autonomous AI. And yes, if an object can have zero velocity, a smart solution can, at least in this blog post, have zero smartness in it.

Although building top-notch AI solutions is, I admit, very satisfying, the complexity level of a system should never be a goal but a means instead. Theoretically, the required intelligence level is mostly defined by the complexity of the surrounding environment, whereas the level of autonomy is mostly a design question. If hard-coded rules provide the needed value, unnecessary complexity is – well – unnecessary.

Unfortunately, choosing the most valuable approach given the task at hand is rarely easy yet important. 

When talking of AI solutions, in many cases more than one field of the table is applicable and the sky’s the limit for development costs. If the outcomes of the two streams are alike, profitability is dictated by the costs. What adds to the difficulty is that solution providers and especially software providers tend to anchor the discussion to some tools or fields favorable for them, although the relevant perspective for the buying companies is the problem. Even when developing the solution in-house, a trusted advisor guiding the way is often invaluable.

It would be very tempting to offer some good-looking checklists or flowcharts for picking the best approach for your very specific business problem. Unfortunately, the devil is in the details, and designing a solution to a unique problem is perhaps even closer to art than science. No matter how agile we are, the importance of good design cannot be overestimated, especially since smart solutions tend to be non-trivial to implement. Consider an example: when building a simple web application, the first wireframe can be put to test and iterative development in a matter of days. By contrast, simply gathering enough training data for a computer vision application, if pre-trained models are not enough, might take weeks meaning the lead time to prove the concept can be remarkably longer. 

To put it another way, a well-used multidisciplinary understanding of the domain at hand, solution possibilities, service design, and sometimes even business design provide the foundations of a successful project.

An illustrative example from the famous Ackoff’s fables is the story of making the already optimized elevators of a high-rise building move perceivably faster by installing mirrors in the waiting lobby and thus shortening the observed waiting time. In other words, humans are not machines, and however biased their observations are, those form their realities. What if your company is in the B2B world and no irrational consumers are involved? Well, we’ve rarely seen software used in a vacuum. 

Even the best, tailor-made AI tool will be worthless if it is not adopted by the relevant users, be they your customers, employees, or even some other stakeholders.

Luckily, adding semantic mirrors to AI-elevators might be as simple as making the machines mimic human reasoning and letting their users hear their inner speech. Another good way to build buy-in is to involve in the development of the people whose judgment is being augmented or replaced. All in all, a well-made smart solution rarely just replaces an existing gearwheel in the abstract money-making machine but transforms whole ways of working instead. Getting buy-in for an itchy-to-use transformation is difficult. The chain is as weak as its weakest link and an AI tool is as useful as its UX.

Concrete analogy: Sensor fusion provides value regardless of levels of autonomy and intelligence

Let us provide some concreteness to this abstract problem by considering tasks in which decisions are to be made based on visual, auditory, or haptic observations and some acts to be conducted according to the decisions. And misusing the term a bit, let us call this vaguely sensor fusion, although a single sensor might be enough. Keeping in mind that the strengths and weaknesses of the four archetypes are nearly similar across smart solutions areas from natural language processing to demand to forecast, let us dive into concrete examples and take a closer look at some interesting sensor fusion applications we at Reaktor have seen in our work.

Infographic of different levels of machine autonomy and intelligence. X-axel: Autonomy. Y-axel: Intelligence. Top left: X-ray + a doctor-computer prediction. Ship in distress giving suggestions. Top right: An autonomous robot. A swarm of autonomous robots working together. Bottom left: AR-goggles. CCTV camera. A machine conducting self-diagnostics. Bottom right: Robots assembling a car. A thermostat.
Top left: X-ray + a doctor-computer prediction. Ship in distress giving suggestions. Top right: An autonomous robot. A swarm of autonomous robots working together. Bottom left: AR-goggles. CCTV camera. A machine conducting self-diagnostics. Bottom right: Robots assembling a car. A thermostat.

 

Data representation: boosting human senses with digital sensors

We’ve been using the simplest form of sensor fusion for ages already: it’s us with displays and cameras. Although cameras or any other sensors don’t do anything alone, by zooming out a bit we can see the potential of strengthening the weakest links of human senses with suitable hardware. Since tools are useless without a user, we prefer talking of human-machine pairs in this quadrant of the table instead of the tools alone. From a bird’s eye view, we humans are unbelievably good at working in uncontrolled and complex environments, combining visual information with auditory and haptic information, and making complex, context-dependent decisions. Our key disadvantages, on the other hand, are that we have a limited field of view and that we tolerate little physical elements. Already using sensors and monitors to let us ingest wider information bandwidth from multiple places allows us to achieve a lot.

Human-machine pairs might seem boring but are surprisingly capable. The same misleading simplicity applies to their demands: merely building a digital tool to meet some technical specifications often leaves a lot of potential unleashed. Just think of the amount of time you have spent marking working hours to an ERP or fixing mistakes caused by the laggy response. Especially when the user of a system is placed to a distance from the actual environment to monitor and control, the requirements e.g. end-to-end connection speeds might be surprisingly high. And as always when a tool is meant to be used more than once, the importance of good service and UX design cannot be overstated. Improvements in UX and service design have been elementary in getting more from well-established technologies.

Hand-held spectral or thermal cameras and remote control facilities are good examples of the benefits of representing data with some predefined logic although they have been introduced a good while ago. However, progress in the area has been anything but halted and modern hardware, e.g. latest AR goggles, enables many useful applications. Examples of such are projecting ultrasound stream together with tool alignment suggestions straight to a doctor’s field of vision, displaying the floor plan of a building to a firefighter’s visor, or providing granular diagnostics from a fleet of manufacturing machines with service suggestions. 

Classic automation: achieving human-defined results with hard-coded machines

Keeping the amount of intelligence constant but letting machines or software do the tasks, we get to classic automation. By classic, we mean the type of automation that has been here for a good while and in which the operation logic combining targets to actions is explicitly defined. The clear difference to plain data representation is that machines excel at repetitive tasks, swift decisions, and tasks otherwise unsuitable for humans. The related technology is mostly tried and tested, and building the solutions is usually fairly straightforward albeit not necessarily easy. Although classic automation might seem simple, it’s surprisingly capable. The autopilot of an airplane is not intelligent but invaluable and a classic automation example instead. If hard-coded rules or explicit logic is enough to complete the tasks, there’s simply no need for higher intelligence or unnecessary complexity.

However, there’s one clear caveat: business operations tend to change over time, and cutting too many corners in the first implementation might make an adaptation of the system to new tasks or environments impossible. In other words, good design and architecture are very important in making the system endure time. And if the human-machine pairs already require fast real-time environments, automated systems do so too, and often even faster ones.

In addition to the autopilots, classic automation examples range from plain radiator thermostats to assembly robots and even whole manufacturing lines. Common for them is that the desired outputs, be they indoor temperatures or welding patterns, need to be changed now and then. If you have ever struggled to get comfy water from a shower faucet (there’s a thermostat inside), you probably agree the UX is not a triviality even in simple-ish solutions.

Augmented intelligence: using software to provide support in complex decisions

Classic automation and augmented intelligence are nearly mirror images of each other. Where the former essentially do easy things fast and alone, the latter helps humans make complex things that might take some time. The key difference to plain data representation then is that the supporting machine does some information processing which can be considered intelligent.

The clear strength of augmented intelligence is using modern tools to reduce the cognitive load of a human by e.g. preprocessing some images while letting the human make the decision. Since humans are prone to biases and have restrictions in decision-making, systems using augmented intelligence emphasize the need for good UX and service design to mitigate the risk of biased decisions. On the other hand, machines do not dominate humans in uncontrolled environments and together they can achieve more than either could alone. For a long time, the winning team in chess was a team of a human and a computer.

The applications of augmented intelligence are many, and the amount of them is constantly increasing in various fields from life sciences to manufacturing and logistics. An interesting example is predicting the seriousness of pneumonia with a machine based on medical imaging and patient data and letting a doctor prioritize the suggested treatments. A very different example yet a fascinating case is using sensors to predict the survivability of a vessel in case of water ingress and giving the officers suggestions on tackling the emergency based on the predictions.

Autonomous AI: letting machines work and learn alone in controlled environments

Intelligent and fully autonomous systems can currently only work on a narrow definition of a task. The environment must be controlled, and surprises must be avoided. It’s the only way we can ‘let’ the system be autonomous. If an environment is controlled enough and volumes or values to handle are high enough, autonomous systems can be immensely valuable. Essentially, they provide the best of the two dimensions and can work at a large scale without human intervention.

However, there are two main challenges with full automation. First, it is not applicable if human judgment is needed either for ethical or for practical reasons. Second, full automation requires the environment to be fairly controlled and predictable. Then again, even in such cases often smaller subprocesses can be automated although the whole process cannot.

A third, more practical challenge with fully autonomous systems is their complexity. Shifting either from classic automation or from augmented systems to it requires a lot more from infrastructure, data acquisition, training, testing, documentation, hardware, etc. Thus, the provided value must be high to make the investment profitable. And sometimes assessing the actual potential might take a good bit of business design.

If the provided value allows, utilizing a fully autonomous system can be immensely valuable. The easiest examples to come up with are those that simply replace humans. However, the more interesting ones are those in which multiple systems work automatically in unison or even as a single digital organism. As a concrete example, automatic warehouse robots are useful but the core of possibilities is reached only when they work as a swarm sharing knowledge and tasks flawlessly. Preventive maintenance cuts downtime already by calling for maintenance and spare parts on its own but the benefits are multiplied if the pieces of equipment learn from each other and if the spare parts supply chain learns the demand. The best of AI in retail is obtained by optimizing the supply chain, workforce, and assortment not in silos but as a whole. Even if the examples might sound unrealistic, the technology is already there and if not the first-mover, the second-mover advantage is up for grabs.

AI-cycle: Smart solutions help to achieve what was previously impossible 

Already exhausted from developing your business and still beside your competitors?

We agree with the piece of advice the Red Queen gave to Alice: merely standing still takes all the running, and that is admittedly tiring. However, what she did not say is that running is not the only way to move. 

If you want to overtake the competition, we suggest you take a look at the possibilities of getting on the wheels of modern AI and automation and making hard operations easier. Instead of the Red Queen, it’s better to discuss your situation with someone not stuck to old ways of working.

At Reaktor, we love to partner with forward-thinking companies and are always happy to have a chat on where we could help with our multidisciplinary approach.

Never miss a post